On Transfer Learning for Naive Brain Computer Interface Users

Motor Imagery (MI) based Brain-Computer Interfaces (BCI) typically require the collection of subject-specific calibration data to build a classifier of motor intent. The BCI users are then trained over multiple online sessions with real-time feedback using the calibrated decoder to acquire MI skills. The subject-specific calibration session is thought to be necessary for accurate MI decoding due to the wide variability in electroencephalogram (EEG) signals across the population. The process of acquiring calibration data is long and tedious and includes training individualized decoding models for each subject. Transfer Learning setups can help circumvent this individualized calibration and decoder training phase by using data acquired from previous subjects. This paper first proposes a geometry-aware deep learning architecture that exploits the spatial similarity of MI neural activity between BCI users. We show the efficacy of the proposed approach by classifying the motor intentions of 18 naive BCI subjects. In a subject-specific setting, our proposed method significantly outperforms classical decoding algorithms. Next, we train the proposed network and skip the subject-specific calibration data to mimic a transfer learning setting. We show that our model architecture achieves similar performance to subject-specific decoders in the transfer learning setting. This finding opens the door to robust BCIs that are readily transferable across subjects without the need for subject-specific calibration and individualized decoding models.